In the information transfer systems of today, speed is a major if not the primary consideration. The faster information can be transferred from one location to another, the better. Because of the need for increased speed performance, a number of techniques have been developed to facilitate and to accelerate data transfer.
Typically, information is transferred from a starting location to a destination by first converting the information from a parallel format to a serialized format. Then, the serialized information is transmitted as a data stream across a transmission medium such a coaxial cable, a twisted pair cable, or a fiber optic cable, to the destination. At the destination, a receiver receives the data stream, and processes it to convert the information back into a parallel format. Thereafter, the parallel information is used for its intended purpose. As a general rule, transferring serialized information is slower than transferring parallel information. This is because rather than sending multiple data units at a time, serialized information is sent one data unit at a time. Because of this, sending serialized information can take substantially longer than sending parallel information. As a result, whenever possible, it is preferable to send information in a parallel rather than a serialized fashion.
To enable multiple data units to be transferred across a transmission medium in parallel, multiple transmission lines may be bundled together to form a multi-link channel. An example of such a setup is shown in FIG. 1, wherein a plurality of transmission lines 106 are bundled and used to connect a multi-link transmitter 102 with a multi-link receiver 104. On each transmission line 106, there is sent a separate data stream, and for each data stream, there is a plurality of serialized data units with each data unit occupying a data period. Because there are multiple transmission lines 106, and because a data unit is sent on each line during each data period, multiple data units are received by the multi-link receiver 104 during each data period. Thus, even though each of the transmission lines 106 carries a serialized data stream, because of the presence of the multiple transmission lines 106, the multi-link receiver 104 is able to receive multiple data units per data period (i.e. receives data units in parallel). By using multiple is transmission lines in this manner, it is possible to transfer information from a starting location to a destination much more quickly.
The multi-channel implementation of FIG. 1 does not come without its problems, however. One of its major problems is the phenomenon known as inter-pair skew. With inter-pair skew, data streams arrive at the multi-link receiver 104 out of alignment. This is so even when the data streams leave the multi-link transmitter 102 completely aligned with each other. Inter-pair skew may be caused by a number of different factors, such as differences in characteristics between the various transmission lines 106, different distances traveled on each transmission line 106, as well as other factors. The inter-pair skew phenomenon is illustrated in the timing diagram of FIG. 2 for the system shown in FIG. 1. Specifically, dt(1) and dt(n) represent two data streams as they leave the transmitter 102 on transmission lines 106(1) and 106(n), respectively, while dr(1) and dr(n) represent the data streams as they arrive at the receiver 104. Notice that the data transition edges of dt(1) and dt(n) are aligned with each other, but the transition edges of dr(1) and dr(n) are not. This shows that while the data streams were aligned with each other at the time they exited the transmitter 102, they are not aligned when they arrive at the receiver 104. This misalignment makes it difficult for the receiver 104 to recover data from both data streams dr(1), dr(n).
To elaborate, a receiver typically recovers data from a data stream by generating a clock signal that is both synchronized and aligned with the data stream. Then, the clock signal is used to generate latching control signals, which are used to latch or sample each of the data units in the data stream. It is usually desirable to generate each latching control signal such that it falls near the middle of a data period so that each data unit is sampled during a time at which its data value is most stable. Since the timing of the latching control signals relative to the data periods of the data stream is important, and since the latching control signals are generated based upon the clock signal, it follows that the alignment of the clock signal relative to the data periods of the data stream is also important. When multiple data streams arrive at a receiver out of alignment, it is impossible for a single clock signal to maintain proper alignment with all of the data streams. As a result, a typical multi-link receiver 104 cannot use the same clock signal to recover data from all data streams dr(1)-dr(n). In general, inter-pair skew makes recovering data from multiple data streams significantly more difficult and complex.
One possible solution to the inter-pair skew problem is to oversample each data stream. That is, during each data period, instead of generating just one latching control signal, multiple latching control signals are generated. These latching control signals are then applied to each data stream to latch in the data units of that data stream. Typically, oversampling is implemented using three, four, or more latching control signals per data period, with each latching control signal being separated from another latching control signal by a certain time constant. By sampling each data unit at multiple points in time, it is likely that at least one of the samples will be taken near the midpoint of the data period where the data unit is stable. By using the data value from that sample, the value of the data unit can be ascertained. Notice that with oversampling, it is not necessary to generate a clock signal that is synchronized and aligned with each data stream. By removing the requirement for a synchronized and aligned clock signal, oversampling solves the inter-pair skew problem. With oversampling, the same set of latching control signals can be applied to all of the incoming data streams.
With reference to FIG. 3, there is shown a timing diagram for a typical implementation of oversampling. Specifically, FIG. 3 depicts two incoming data streams dr(1) and dr(n) which are out of alignment with each other. In addition, FIG. 3 shows four latching control signals C1, C2, C3, and C4, all of which trigger within the span of one data period. For the sake of simplicity, only four control signals C1-C4 are shown; however, it should be understood that there are four control signals per data period. The control signals C1-C4 within a data period are separated from each other by a time constant TC. Depending upon the relative alignment between each data stream dr(1), dr(n) and the control signals C1-C4, one of the controls signals C1-C4 will be used as the preferred signal for extracting data from the data stream. For example, it appears from FIG. 3 that latching control signal C2 falls closest to the midpoint of the data periods for data stream dr(1); thus, signal C2 will be the preferred signal for extracting information from data stream dr(1). In contrast, for data stream dr(n), it appears that control signal C3 falls closest to the midpoint of the data periods. Consequently, signal C3 will be the preferred signal for extracting data from data stream dr(n). By generating multiple latching control signals per data period, it is possible to use the same set of latching control signals to extract data from multiple data streams, even when those data streams are misaligned relative to each other.
Oversampling provides a solution to the inter-pair skew problem. However, this solution does not come without significant drawbacks. For one thing, generating and managing multiple latching control signals per data period imposes significant additional overhead, both in terms of generating additional signals and of providing additional storage (since each sampled value needs to be stored). As a result, there is significant redundancy in an oversampling system. Also, notice that the sampling position is not optimal. Rather than sampling a data unit at the midpoint of a data period, oversampling samples the data unit at positions which are just somewhat close to the midpoint. This may lead to data errors. In addition and perhaps most significantly, oversampling imposes a practical limit on how quickly data can be transmitted. By requiring a small delay (time constant TC) between successive latching control signals, oversampling limits the rate at which data can be practically sent in a data stream.
To elaborate, a delay is typically imposed on a signal by passing the signal through a delay element. The delay provided by the delay element depends upon how quickly the element processes the signal. As a general rule, the faster an element processes a signal (and hence, the shorter the delay), the more difficult and expensive the element is to manufacture. Due to constraints such as cost and technological limitations, there is a practical limit on how quickly a delay element can process a signal, and hence, how small a delay can be. For purposes of discussion, this limit will be referred to as TCmin and it will be assumed that the time constant TC (FIG. 3) between the control signals C1-C4 is set to TCmin. As noted above, there are typically four latching control signals per data period in an oversampling implementation; hence, there are four TCmin's per data period. Since TCmin is a minimum, this means that the shortest data period that a data unit can have is four times TCmin. Since frequency is the inverse of period, this in turn means that the highest frequency that a data stream can have is ¼TCmin. Since data rate is directly proportional to frequency, it follows that data rate is also limited by the factor ¼TCmin. As this discussion shows, TCmin plays a large role in limiting the rate at which data can be sent in a data stream. Depending upon the magnitude of TCmin, is the maximum achievable data rate may be much lower than would be desired. This aspect of oversampling makes it an unattractive solution in many implementations.
As an alternative to oversampling, a plurality of phase locked loops (PLL) may be used to recover data from the multiple data streams. According to this implementation, an individual PLL is coupled to receive each of the data streams. Each PLL generates a clock signal that is synchronized and aligned with the data stream that it receives. Then, based upon that clock signal, the PLL generates a plurality of latching control signals, one for each data period of the data stream. Each latching control signal falls around the midpoint of each data period. Once generated, the latching control signals are used to each of the data units of the data stream. In a PLL implementation, each PLL works independently to recover data from the data stream with which it is associated. Because data from each data stream is recovered independently, misalignment among the various data streams has no affect on the operation of the system. Thus, even in the face of inter-pair skew, a PLL implementation is able to recover data from multiple data streams.
While the PLL implementation provides a possible solution to the inter-pair skew problem, it comes with some significant drawbacks. First, because it uses PLL's, jitter performance is less than optimal, and power consumption is relatively high due to the nature of PLL's. In addition and more significantly, the PLL implementation, like oversampling, can significantly limit the rate at which data can be transferred in a data stream. To illustrate this problem, reference will be made to FIG. 4, wherein a timing diagram for one of the PLL's in a typical PLL implementation is shown. For the sake of simplicity, only one latching control signal C1 is depicted in FIG. 4; however, it should be understood that there is a latching control signal for each data period.
As shown in FIG. 4, the control signal C1 is generated by delaying the Clock signal by a time constant TC, which is one half of a data period. Since TC is one half of a data period in duration, it follows that there are two TC's per data period. Assuming that TC is set to the minimum practical value TCmin as discussed above, it means that the minimum data period that a data stream can have is 2TCmin. Given that frequency is the inverse of period, and that data rate is directly proportional to frequency, if follows that the highest data rate achievable using the PLL implementation is limited by the factor ½TCmin. Thus, even though it is not as limiting as the oversampling implementation, the PLL implementation still does impose a significant limit on the rate at which data can be sent in a data stream. Given all of the drawbacks of the current implementations discussed above, there is a need for an improved mechanism for concurrently processing multiple data streams.